iris recognition under various degradation models

08.06.2007 1

Iris Recognition Iris Recognition Under Various Under Various

Degradation ModelsDegradation Models

Hans Christian SagbakkenHans Christian Sagbakken

08.06.2007 2Hans Christian Sagbakken

OutlineOutline

IntroductionIntroduction Scope and research Scope and research

questionsquestions Experimental setupExperimental setup Results Results ConclusionsConclusions


IntroductionIntroduction


Biometrics technologyBiometrics technology

Biometrics refers to technologies that Biometrics refers to technologies that measure and analyze human physical measure and analyze human physical and behavioural characteristics and behavioural characteristics

Examples of characteristics include Examples of characteristics include fingerprints, eye retinas and irises, fingerprints, eye retinas and irises, facial patterns and handfacial patterns and hand measurementsmeasurements

Two main application:Two main application: Verification (mobile banking) Verification (mobile banking) Identification (security control)Identification (security control)


Example of an iris Example of an iris patternpattern


Iris recognition prosessIris recognition prosess1. Segmentation prosess

2. Normalisation prosess

3. Iris code generation

Comparison

4. Comparison/decision


Scope and research Scope and research questionsquestions


Research questionsResearch questions

1. Under which conditions is iris-1. Under which conditions is iris-based recognition feasible?based recognition feasible?

2. Which filter to perform under 2. Which filter to perform under certain degradation conditions?certain degradation conditions?


Scope of the thesisScope of the thesis

The thesis was restricted to experiments The thesis was restricted to experiments in MATLAB onlyin MATLAB only

Adapt Libor Masek’s open source code Adapt Libor Masek’s open source code for the experiments (different filters, for the experiments (different filters, inter-class and intra-class comparisions)inter-class and intra-class comparisions)

The iris images degradations are The iris images degradations are simulated in MATLAB with different simulated in MATLAB with different parameters (to find the best filter under parameters (to find the best filter under different conditions)different conditions)


Experimental setupExperimental setup


ImplementationImplementation

Expanded Libor Masek’s open source Expanded Libor Masek’s open source code for iris recognition with four code for iris recognition with four filtersfilters

Log-Gabor filter (9600 bit, original filter)Log-Gabor filter (9600 bit, original filter) 702-bit Haar wavelet filter702-bit Haar wavelet filter 87-bit Haar wavelet filter87-bit Haar wavelet filter Log of Gaussian filter (9600 bit)Log of Gaussian filter (9600 bit)

Expanded the search function with Expanded the search function with inter-class and intra-class inter-class and intra-class comparisonscomparisons


Iris databaseIris database The filters are tested on 500 images from the The filters are tested on 500 images from the

UBiris database. Five images per person for UBiris database. Five images per person for 100 persons.100 persons.

The images are simulated with different The images are simulated with different paramentersparamenters Add noise in the image database (Gaussian noise) Add noise in the image database (Gaussian noise) Add blur in the image databaseAdd blur in the image database Change the light intensity in the image databaseChange the light intensity in the image database Rotate the images in the databaseRotate the images in the database


EvaluationEvaluation For each filter under different conditions, the For each filter under different conditions, the

False Acceptance Rate (FAR) and False False Acceptance Rate (FAR) and False Rejection Rate (FRR) are computedRejection Rate (FRR) are computed

Inter-class comparisons (to experiment with Inter-class comparisons (to experiment with FAR). For each test 123,750 comparisons are FAR). For each test 123,750 comparisons are donedone

Intra-class comparisons (to experiment with Intra-class comparisons (to experiment with FRR). For each test 1000 comparisons are FRR). For each test 1000 comparisons are donedone

Totally 6,930,000 inter-class and 56,000 intra-Totally 6,930,000 inter-class and 56,000 intra-class comparisons are performed.class comparisons are performed.


Example of hamming Example of hamming distributiondistribution

Inter-class comparisons

Intra-class comparisons


Example of FAR and FRRExample of FAR and FRR

Optimal threshold value = 0.32


ResultsResults


Results under noisy Results under noisy conditionsconditions

Støyvarianse 0.002 Støyvarianse 0.004 Støyvarianse 0.006

Terskel FRR FAR Terskel FRR FAR Terskel FRR FAR

702-bit Haar wavelet 0.33 0.265 0.189 0.35 0.262 0.249 0.36 0.315 0.302

87-bit Haar wavelet 0.42 0.304 0.298 0.42 0.397 0.291 0.43 0.392 0.344

Log-Gabor 0.42 0.280 0.196 0.43 0.412 0.277 0.44 0.528 0.295

Log of Gaussian 0.39 0.259 0.191 0.41 0.300 0.287 0.42 0.384 0.336


Results under blur Results under blur conditionsconditions

Blur radius 2 Blur radius 4 Blur radius 6


702-bit Haar wavelet 0.30 0.126 0.118 0.31 0.134 0.122 0.34 0.221 0.180

87-bit Haar wvelet 0.37 0.181 0.128 0.40 0.178 0.154 0.43 0.216 0.210

Log-Gabor 0.39 0.127 0.109 0.39 0.138 0.128 0.41 0.263 0.195

Log of Gaussian 0.30 0.137 0.135 0.27 0.155 0.148 0.25 0.251 0.223


Results under light Results under light changeschanges

Lysintensitet -10% Lysintensitet -5% Lysintensitet +5% Lysintensitet +10%

Terskel FRR FAR Terskel FRR FAR Terskel FRR FAR Terskel FRR FAR

702-bit Haar wavelet 0.26 0.225 0.172 0.29 0.175 0.159 0.32 0.129 0.128 0.33 0.130 0.113

87-bit Haar wvelet 0.37 0.243 0.207 0.41 0.228 0.229 0.43 0.230 0.224 0.43 0.256 0.217

Log-Gabor 0.37 0.205 0.198 0.39 0.168 0.127 0.42 0.125 0.120 0.43 0.146 0.122

Log of Gaussian 0.31 0.212 0.176 0.31 0.198 0.156 0.26 0.251 0.210 0.33 0.153 0.126


Results under rotationResults under rotation

2 grader 3 grader 4 grader


702-bit Haar wavelet 0.35 0.230 0.209 0.32 0.179 0.151 0.35 0.229 0.213

87-bit Haar wavelet 0.40 0.195 0.230 0.40 0.269 0.187 0.42 0.281 0.259

Log-Gabor 0.40 0.148 0.139 0.40 0.160 0.135 0.42 0.206 0.174

Log of Gaussian 0.34 0.178 0.147 0.35 0.173 0.172 0.37 0.277 0.232


ConclusionsConclusions Under noisy conditions the best results where Under noisy conditions the best results where

achieved with 702-bit Haar wavelet filterachieved with 702-bit Haar wavelet filter Under blur conditions the best results where Under blur conditions the best results where

achieved with 702-bit Haar wavelet filterachieved with 702-bit Haar wavelet filter Under light changes the best results where Under light changes the best results where

achieved with 702-bit Haar wavelet filter and achieved with 702-bit Haar wavelet filter and Log-Gabor filterLog-Gabor filter

Under rotation the best results where achieved Under rotation the best results where achieved with Log-Gabor filterwith Log-Gabor filter

Totally the best filter is 702-bit Haar wavelet Totally the best filter is 702-bit Haar wavelet filterfilter


Questions???Questions???

iris recognition under various degradation models

Documents

best filter

doneintraclass comparisons

best results

haar wavelet filterunder

haar wavelet filter87

filtersloggabor filter

experiments different

different paramenters